On Wed, Aug 26, 2015 at 12:10:42PM -0500, Rich Johnston wrote: > On 08/26/2015 11:42 AM, Eric Sandeen wrote: > >On 8/26/15 9:33 AM, rjohnston@xxxxxxx wrote: > >>The call to memset will segfault because the offset for the first > >>parameter is done twice. We are using pointer math to do the > >>calculation. > >>The first time is when calculating oldsize, the size of i2gseg_t > >>is accounted for. > >> oldsize = (numsegs - SEGPERHNK) * sizeof(i2gseg_t); > >>Then in the call to memset, oldsize is again multiplied by the size > >>of i2gmap_t. > >> memset(inomap.i2gmap + oldsize, ...) > >> > >>i2gmap holds the used inodes in each chunk. When there are 2^31 chunk > >>entries, it could describe 2^31 (1 inode/chunk)- 2^40 (64 inodes/chunk). > >> > >>With 100s of millions of inodes there are enough entries to wrap the > >>32 bit variable oldsize. > >> > >>Switching to use array index notation instead of calculating the > >>pointer address twice ;) would resolve this issue. The unneeded > >>local variable oldsize can be removed as well. > > > >Ok, this doesn't explain the type change for numsegs, does it? > Nope that was a typo, as it is an array index I meant to change it from > a signed int (intgen_t) to an unsigned (uint32_t). That doesn't fix anything. If you push the index down through zero by decrementing it too much, it will still result in an array bound overrun. i.e. It will index array element UINT_MAX rather than -1. And the compiler still won't catch it because it can't bounds check runtime calculated index values... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
